Infrastructure for Parallel Processing Research

该奖项为开发并行语言和编译器、算法的自动并行化、计算机视觉以及数字设备的模拟和合成等领域的计算基础设施提供支持。基础设施由最先进的单指令多数据(SIMD)计算机和支持人员组成，以确保计算机硬件和软件得到适当维护。并行计算机的并行语言和编译器的研究集中在高效软件对底层并行体系结构的依赖问题上。这种依赖使得开发同样高效的可移植软件变得困难，而且到目前为止也是不可能的。研究的重点是可以在不依赖体系结构假设的情况下表达并行性的语言，以及将使用这些无体系结构构造的程序编译成针对特定并行体系结构的高效代码的编译器。算法自动并行化研究的重点是如何将算法高效地映射到SIMD计算机上。SIMD计算机通过让多个处理器对不同的数据执行相同的指令来执行。因此，数据的移动和选择哪些处理器应该执行指令对于获得高性能是至关重要的。这项研究的重点是一个软件工具，它可以适应各种SIMD架构，并将执行用于这些架构的算法的并行化。计算机视觉和图像处理一直是SIMD计算机的理想应用，因为图像的处理可以分解为子图像的处理，每个子图像的算法是相同的，然后在处理器之间进行统一的数据移动。计算机视觉研究旨在开发一套高效使用SIMD架构的视觉和图像处理算法。数字设备的模拟和合成研究项目使用SIMD计算机的处理能力来执行需要超级计算机计算能力的任务。这些项目包括数字衍射元件的合成和半导体器件操作的模拟。用于光学的数字衍射元件的合成需要相当大的计算能力和输入/输出能力(在VAX780上大约为10,000小时)。同样，半导体器件的建模需要大量的计算资源(求解30,000阶稀疏矩阵)。这两个研究项目都建议开发在SIMD计算机上使用的算法，以便获得解决这种规模的问题所需的计算能力。//

This award provides support to develop computing infrastructure in the areas of parallel languages and compilers, automatic parallelization of algorithms, computer vision, and simulation and synthesis of digital devices. The infrastructure consists of a state of the art Single Instruction Multiple Data (SIMD) computer along with support staff to ensure that the computer hardware and software are properly maintained. The parallel languages and compilers for parallel computers research focusses on the problem of the dependence of efficient software on the underlying parallel architecture. This dependence makes the development of portable software which is also efficient difficult and to date impossible. The research focusses on languages in which parallelism can be expressed free of architecture assumptions and compilers that will compile programs using these architecture free constructs into efficient code for specific parallel architectures. The automatic parallelization of algorithms research focusses on the problem of efficiently mapping algorithms to SIMD computers. The SIMD computer executes by having many processors execute the same instruction on different data. Thus the movement of data and the choice of which processors should execute an instruction is critical in order to obtain high performance. This research focusses on a software tool that can be adapted to a variety of SIMD architectures and which will perform the parallelization of algorithms to be used with these architectures. Computer vision and image processing have been ideal applications for SIMD computers because the processing of images can be broken down into the processing of sub-images, each sub-image algorithm being the same, followed by uniform data movement between processors. The computer vision research aims to develop a collection of algorithms for vision and image processing that will efficiently use the SIMD architecture. The research projects on the simulation and synthesis of digital devices use the processing power of the SIMD computer to perform tasks that would require supercomputer computational capabilities. These projects include the synthesis of digital diffractive elements and the simulation of the operation of semiconductor devices. The synthesis of digital diffractive elements, used in optics, requires substantial computational capabilities and input/output capacity (on the order of 10,000 hours on a VAX 780). Similarly the modeling of semiconductor devices requires substantial computational resources(the solution of order 30,000 sparse matrices). Both of these research projects propose to develop algorithms for use on the SIMD computer in order to obtain the computational capability necessary to solve problems of this magnitude. //